Weekly AI Roundup: What Happened in AI This Past Week (July 25–31, 2026)

Welcome back. Late July in Dhaka is a study in saturation — the monsoon has settled into its rhythm, the lanes hold their puddles a little longer each evening, and the air sits heavy over the rooftops from Mirpur to Agrabad. The AI world, meanwhile, finished counting its money. The last and largest of the earnings reports landed this week — Microsoft, Meta and Amazon, three of the companies whose data centres carry most of the world’s AI — and the numbers were of a size that resists intuition. In the same seven days, something stranger and possibly more important happened: more than a thousand of the people who build these systems signed their names to a public letter asking their own government for the tools to slow the technology down if that ever becomes necessary. And in Brussels, the world’s first comprehensive AI law reached the eve of its enforcement date.

Two halves as always. Five global, five from home. The same two questions on each — who benefits, who pays — and the same effort to read every story from Dhaka as well as from Redmond, Seattle or Brussels. And, as ever, the discipline of checking each load-bearing fact against its source before repeating it. This was a week of enormous numbers, and enormous numbers deserve extra suspicion. Let us begin.

5 Big International Developments

1. Microsoft’s Azure crossed $100 billion a year — and grew faster doing it

The biggest of the week’s ledgers opened in Redmond. On July 29, Microsoft reported fourth-quarter revenue of $90.0 billion, up 18 per cent year on year, with Azure and other cloud services growing 43 per cent — and confirmed that Azure surpassed $100 billion in annual revenue for the first time in its 2026 fiscal year, while Microsoft 365 Copilot passed 30 million paid seats. Full-year revenue came to $331.8 billion, up 18 per cent, and the company’s commercial remaining performance obligation — contracted work not yet delivered — swelled 84 per cent to $678 billion. Chief executive Satya Nadella said: “We are advancing the frontier on the cost-to-outcome curve, ensuring every customer can turn tokens into business results,” adding that the Azure and Copilot milestones reflect “the confidence customers are placing in us to power their AI transformation.” The quarter also included a $3.2 billion gain on Microsoft’s investment in Anthropic — part of discrete items adding $0.27 to earnings per share — and adjusted earnings of $4.74 beat expectations near $4.24, per CNBC; every other figure comes from Microsoft’s own release. (A plain-language note: a “remaining performance obligation” is the backlog — money customers have already contracted to spend that the company has not yet delivered against. It is a rough measure of how much future demand is signed and waiting.)

The significance is acceleration at scale. A business the size of Azure is not supposed to speed up — yet cloud growth of 43 per cent, on a base that just crossed $100 billion a year, says enterprise AI demand is still compounding rather than plateauing. The $678 billion backlog says the same in a different register: years of AI compute already committed before a single new server is racked.

Who benefits, and who pays? Microsoft and its shareholders benefit most obviously, as do the enterprises whose AI workloads now run reliably at scale. The ones who pay are Microsoft’s competitors, who must match this pace of investment or concede the enterprise market — and, potentially, customers in markets with little bargaining power, if cloud concentration hardens into pricing power.

The Dhaka view is the one this newsletter keeps returning to: Bangladesh will rent this infrastructure, not build it. When the world’s AI runs on a handful of clouds, the terms of access — pricing, data residency, availability zones — become questions of national digital-economy policy. The practical work is negotiating those terms well and building the skills layer that stays valuable no matter whose cloud wins.

2. Meta grew 28 per cent — and spent so hard its free cash flow nearly vanished

The week’s most unnerving report came from Menlo Park. On July 29, Meta reported second-quarter revenue of $60.8 billion, up 28 per cent year on year — but total costs jumped 55 per cent to $42.0 billion, operating income fell 8 per cent, net income fell 14 per cent to $15.9 billion, and free cash flow collapsed to $784 million from $8.55 billion a year earlier, as quarterly capital spending reached $31.1 billion. The company narrowed its 2026 capital-expenditure outlook to $130–145 billion — raising the bottom of the range — and guided full-year expenses to $165–169 billion. The quarter also carried $2.4 billion in legal charges and $1.18 billion in severance from a headcount reduction announced in May. Founder and chief executive Mark Zuckerberg said: “AI is accelerating our core business today, powering our next generation of products, and opening the door to entirely new enterprise opportunities. The results are already showing, and I’m optimistic about the potential ahead.” Investors were less sure: the shares fell about 10 per cent after the report, according to CNBC. All financial figures come from Meta’s own results release. (A plain-language note: “free cash flow” is the cash a business generates after paying for its operations and its capital spending — the money actually left over after the building is paid for.)

The significance is that Meta has become the market’s referendum on AI capital spending. Revenue growing 28 per cent is an excellent business by any standard; costs growing 55 per cent is a bet that swallows it. A $130–145 billion single-year buildout asks investors to accept years of near-zero cash generation on faith that the AI products arrive.

Who benefits, and who pays? If the bet lands, Meta’s advertisers, its billions of users and its shareholders eventually benefit from better AI woven through its products. Right now, shareholders are the ones visibly paying — in evaporated cash flow and a 10 per cent haircut — along with the roughly 8,000 employees cut in May as spending shifted from people to data centres.

The Dhaka view is closer to home than it looks. For millions of Bangladeshis, Facebook effectively is the internet — the shop window, the news feed, the family album. The company betting its cash on AI means the products Bangladesh’s small businesses depend on will change quickly, and largely without consultation. The lesson for local firms is Meta’s own, inverted: ambition needs a ledger. Spend on what compounds; measure ruthlessly.

3. Amazon’s AWS grew its fastest in 18 quarters — and its chip business passed $25 billion

The third ledger belonged to Seattle. On July 30, Amazon reported second-quarter net sales of $200.6 billion, up 20 per cent year on year, with AWS growing 37 per cent to $42.2 billion — its fastest growth in 18 quarters — and said its AI business and its chips business had each passed annualized run rates of more than $25 billion. Chief executive Andy Jassy put it plainly: “AWS is booming, growing 36.7% year-over-year in Q2—our fastest growth in 18 quarters—and our AI and Chips businesses each eclipsed run rates of more than $25 billion.” Two asterisks belong on the quarter’s eye-watering $62.6 billion net income: $53.4 billion of it was a non-operating gain “primarily from our investments in Anthropic” — paper wealth, not sales — and the company’s trailing-twelve-month free cash flow was actually negative by $7.6 billion, as purchases of property and equipment rose $66.1 billion year on year, primarily for AI. Amazon said Anthropic and OpenAI have made “multi-year, multi-gigawatt” commitments to its Trainium chips. On the earnings call, per CNBC, the company guided 2026 capital spending to roughly $220 billion, citing rising memory costs. The press-release figures come from Amazon’s own filing. (A plain-language note: an “annualized run rate” takes the latest period’s revenue and multiplies it out to a full year — a speedometer reading, not a distance travelled. “Trainium” is Amazon’s in-house AI chip family, built so its cloud is less dependent on Nvidia.)

The significance is twofold. AWS reaccelerating from 17 per cent a year ago to 37 per cent ends the narrative that Amazon was losing the AI cloud race. And a custom-chip business crossing $25 billion — with the two most prominent AI labs signing multi-gigawatt commitments — is the first proof at scale that there is a credible alternative to Nvidia for training frontier models.

Who benefits, and who pays? Cloud customers benefit from a three-way fight between Microsoft, Google and Amazon, and chip buyers benefit from real competition in AI silicon. The ones who pay, for now, are Amazon’s own cash reserves — negative free cash flow on $200 billion of quarterly sales is a choice — and eventually anyone exposed if $220 billion of annual spending meets a demand curve that bends.

The Dhaka view is that competition upstream is the best news a downstream country gets. Bangladesh cannot influence chip markets, but every credible rival to Nvidia and every price war among clouds lowers the cost floor for a Dhaka startup renting inference by the token. Watch the price lists, not the poetry.

4. More than 1,300 frontier-lab researchers asked Washington for the tools to slow AI down

The week’s most remarkable document was not an earnings release. In an open letter published on July 28, verified employees of frontier AI companies — more than 1,300 of them by Friday, on the letter’s live counter — asked the US government to “support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development.” The signatories are not outsiders: they include Anthropic chief executive Dario Amodei, OpenAI chief scientist Jakub Pachocki and chief research officer Mark Chen, Google DeepMind co-founder and chief AGI scientist Shane Legg, Meta AI chief scientist Shengjia Zhao, Safe Superintelligence chief executive Ilya Sutskever and Thinking Machines chief scientist John Schulman. Meta’s Zhao put the reasoning plainly: “AI is progressing at a rate that our society might not be ready for. Frontier labs are very close to AI that can exceed even the best people on almost every metric of intelligence.” The signatory list and quotes come from the letter’s own site; the July 28 publication date is from press coverage, as the site itself dates it simply “July 2026”. Claims that any company endorsed the letter corporately circulated only in aggregator blogs, so I do not repeat them. (A plain-language note: “automated AI development” means AI systems doing AI research themselves — writing the code and running the experiments that produce the next, stronger system. “Pacing” means having the ability to deliberately control how fast that loop runs.)

The significance is who signed, and what they did not ask for. This is not a demand to pause anything today; it is sitting chief scientists of nearly every major American lab saying, in public, that the machinery for slowing down does not exist and should be built before it is needed. Coming the same week their employers’ landlords reported $100 billion clouds and $220 billion spending plans, the letter reads like the industry’s other ledger — the one denominated in trust.

Who benefits, and who pays? Everyone benefits if verifiable pacing tools exist and are never needed; the cost of building them is trivial against the sums in the stories above. The ones who might pay are the labs themselves, if coordination hardens into regulation they dislike — which is precisely why researchers asking for it voluntarily is notable.

The Dhaka view is that countries like Bangladesh have the most at stake in whether the frontier is paced and the least say in it. If automated AI development accelerates beyond anyone’s control, rule-takers feel the consequences first and last. An international mechanism — with seats for the Global South, at whichever of the world’s proliferating governance tables it is built — is the version of this letter Dhaka should be arguing for.

5. The EU AI Act’s main rules become enforceable on August 2 — the day after this window closes

The countdown this newsletter flagged in mid-July ran out this week. According to the European Commission’s own implementation timeline, on August 2, 2026 “the majority of rules of the AI Act come into force and enforcement starts”: obligations for high-risk AI systems listed in Annex III begin to apply, transparency rules take effect, every member state is to have at least one AI regulatory sandbox, and enforcement begins at both national and EU level. For general-purpose AI models — the layer where OpenAI, Google, Anthropic and Meta live — the obligations have applied since August 2025; what changes now is that the Commission’s power to fine becomes active, with penalties for general-purpose-model providers of up to 3 per cent of global annual turnover or €15 million, whichever is higher, under Article 101 of the regulation as read by multiple legal analyses. One honest caveat, visible on the Commission’s own page: under the proposed “Digital Omnibus” package, the Commission has suggested linking the application of the high-risk rules to the availability of harmonised standards — but that is a proposal, not adopted law, and the official timeline still shows August 2. (A plain-language note: “high-risk” systems under Annex III are AI used in sensitive domains — hiring, credit, essential services, law enforcement and the like — which now carry documentation, oversight and quality obligations. A “regulatory sandbox” is a supervised space where firms can test AI systems with the regulator watching.)

The significance is that the world’s first comprehensive AI law stops being a schedule and starts being a fact. Two weeks ago this newsletter wrote that the Act’s teeth were “weeks away”; they are now days away, and every frontier lab selling into Europe faces its first compliance deadline with financial consequences attached.

Who benefits, and who pays? Europeans subject to AI decisions benefit from enforceable rights; regulators everywhere benefit from watching the first real enforcement experiment. The ones who pay are the companies now carrying genuine legal risk in a major market — and possibly European startups, if compliance costs weigh heavier on small firms than large ones, as critics have long warned.

The Dhaka view is the Brussels effect, again. Bangladeshi software and services firms with European clients will find AI-Act clauses flowing down their contracts — documentation demands, transparency attestations, audit rights — regardless of what Bangladeshi law says. The firms that treat August 2 as their own deadline, not just Europe’s, will be the ones that keep the contracts.

5 Stories That Hit Closer to Home (Bangladesh)

1. The Prime Minister opened Bangladesh’s first national semiconductor summit — with eight roadmaps and real tax measures

The announcement this newsletter covered last week became an event this week. On July 25, Prime Minister Tarique Rahman inaugurated the National Semiconductor Symposium and BEAR Summit 2026 at Dhaka’s Novo Theatre — the country’s first national deep-tech summit, themed “Building Bangladesh’s Semiconductor Future through Global Partnerships” — where eight strategic national documents were unveiled across its sessions, including a Talent Roadmap 2030, an Ecosystem Roadmap and an International Partnership Framework. The Prime Minister said the government had withdrawn all taxes and duties on semiconductor raw materials in the current budget, announced plans for cash incentives on chip-design export earnings and first-ever startup grants of Tk5 crore a year, and declared, in remarks translated from Bangla: “The semiconductor industry could emerge as Bangladesh’s next promising economic sector after the RMG industry.” The two-day gathering — jointly organised by the Bangladesh Semiconductor Industry Association, the science and technology ministry and the Silicon River Bangladesh programme, BEAR standing for Bangladesh Electronics, AI and Robotics per the state news agency — drew Texas Instruments’ chief technology officer Ahmed Bahai as keynote speaker, alongside GlobalFoundries, NXP, Applied Materials, SanDisk and Purdue University, whose Professor Muhammad Mustafa Hussain is Silicon River’s chief architect. Telecom and ICT Minister Faqir Mahbub Anam closed with a line worth keeping: “So this is not just a conference; it is the beginning of a movement.” One small discrepancy, noted honestly: most coverage dates the summit July 25–26; one outlet’s closing report says July 27. (A plain-language note: “deep tech” is shorthand for businesses built on hard science and engineering — chips, biotech, robotics — as opposed to apps built on existing platforms.)

The significance is that the design-first instinct this newsletter has praised for weeks now has paper and money behind it: written national roadmaps, duty withdrawals, export incentives and a head of government personally framing chips as the successor to garments. A summit built around talent, design and partnerships is aimed at the affordable end of the value chain — the only end, as last week’s Intel story showed, that a country without trillion-taka fabs can play.

Who benefits, and who pays? Students, engineers and the small cluster of local chip-design firms benefit from a national convening point, fiscal incentives and a diaspora-linked talent pipeline. The state pays in foregone duties and grants — a modest, sensible wager. The risk of paying more is the perennial one: eight roadmaps is eight documents, and documents do not tape out chips.

The Dhaka view is warmly encouraging, with one raised eyebrow. Last week’s press conference promised a National AI Institute alongside this summit; this week’s summit coverage did not mention it, and I could find no further detail. Roadmaps with budgets become industries; roadmaps without budgets become anniversaries. The Talent Roadmap 2030 is the one to watch — people, not press releases, are the part of the chip value chain Bangladesh can actually own.

2. Bangladesh Bank opened the door to PayPal-style cross-border payments

The central bank had a remarkable week, and this was its centrepiece. In a circular issued July 29, Bangladesh Bank created a framework allowing authorised dealer banks to partner with foreign payment service providers, digital platforms and online payment gateways — collectively termed CDPSPs, cross-border digital payment service providers — with customer “Digital Value Accounts” operating strictly as sub-accounts linked to a bank-held master settlement account. Permitted uses include travel quotas, visa fees, hotel bookings, memberships, IT-related services and small-value international online payments of up to $300 per transaction; balances can draw on exporters’ retention quota and resident foreign-currency accounts, and foreign visitors gain a route to pay local merchants. Banks need prior acknowledgement from the central bank’s Foreign Exchange Policy Department before launching, with real-time integration, mirror ledgers and full know-your-customer checks. The details come from the state news agency’s and The Business Standard’s reporting of the circular; the suggestion that this “could encourage global payment platforms such as PayPal, Payoneer and Stripe to enter the Bangladesh market” is the framing of market observers quoted in that coverage — no platform has announced entry, and I state that plainly. (A plain-language note: an “authorised dealer” is a bank licensed to handle foreign exchange; a “Digital Value Account” is the wallet balance a customer would hold with a payment platform, here required to sit under a bank’s supervision.)

The significance is structural. For over a decade, “why can’t we have PayPal?” has been the single most repeated question in Bangladesh’s freelancer economy, and the honest answer was that no regulatory door existed. This circular builds the door. It does not summon anyone through it — but frameworks precede entrances, and last week’s freelancer forex easing plus this makes a deliberate sequence.

Who benefits, and who pays? Freelancers, small exporters and travelling professionals benefit if global platforms actually arrive; banks gain a new partnership business. The ones who pay are the informal channels that thrived on the old friction — and, if implementation is clumsy, users stuck between a circular’s promise and a bank’s caution.

The Dhaka view is that this is the second week running that the most useful AI-economy news from Bangladesh is about payments, not models. A Dhaka freelancer selling prompt-engineering or annotation services is an AI worker; getting her paid smoothly is AI policy. The plumbing is the strategy.

3. A consolidated export rulebook arrived — with the first dedicated chapter for freelancers

The central bank was not finished. On July 30, Bangladesh Bank issued a revised consolidated export-trade circular — sixteen parts, effective until July 2027 — introducing guidelines for electronic presentation of export documents, a pilot framework for digitally processing trade documents under documentary collections and letters of credit, and, for the first time, a dedicated chapter on foreign-exchange transactions for freelancers and individual service exporters. The new chapter institutionalises the easings reported last week: payment against electronic evidence, remittances up to $20,000 without formal declaration, gateway payments up to $10,000 per transaction, dual-currency cards and a 50 per cent retention quota for ICT freelancers. A caution on sourcing: the July 30 circular was reported by The Business Standard, and I could not find a second outlet’s coverage within the week — the underlying details are consistent with the central bank’s documented pattern, but treat the specifics as single-sourced until the circular text circulates. (A plain-language note: a “letter of credit” is a bank’s promise to pay an exporter once documents prove the goods were shipped; “documentary collection” is the cheaper cousin where banks pass documents without guaranteeing payment.)

The significance is consolidation. Scattered ad-hoc circulars are how well-meaning reforms die — misapplied at branch counters by officers who never saw them. A consolidated rulebook with freelancers as a named chapter turns last week’s concessions into standing policy that a bank officer can be pointed to.

Who benefits, and who pays? Exporters and freelancers benefit from rules they can cite; banks benefit from clarity. The cost, if any, falls on the transition itself — sixteen parts of new guidance only work if banks train for them, and business reaction quoted in the coverage stressed exactly that: implementation, infrastructure and capacity.

The Dhaka view: three interlocking reforms in nine days — freelancer easing, a payments framework, a consolidated rulebook — looks less like coincidence and more like a central bank deliberately rebuilding the export plumbing ahead of LDC graduation. It is the most coherent stretch of digital-economy policymaking this newsletter has had the pleasure of covering. The follow-through test now belongs to the branch counter, not the head office.

4. EBL and bKash put a bank fixed deposit inside a mobile wallet

The week’s fintech story was small, concrete and quietly significant. On July 30, Eastern Bank and bKash announced what they describe as Bangladesh’s first bank-led digital fixed-deposit service inside a mobile financial services app — letting bKash customers open and manage an EBL fixed deposit entirely from their phones, with no branch visit. EBL’s deputy managing director M Khorshed Anowar said: “After the successful launch of EBL DPS on the bKash app, we took the initiative to become the first bank to offer fixed deposits through the platform.” bKash’s chief commercial officer Ali Ahmmed added: “By leveraging bKash’s extensive customer base and digital infrastructure, banks can extend their services to millions of people with greater convenience.” Honesty about sourcing: the coverage is press-release-based and the “first” claim is the companies’ own; rates, tenors and a launch date were not reported. (A plain-language note: a “fixed deposit” locks savings for a set term at a set interest rate; “MFS” — mobile financial services — is the bKash/Nagad layer through which much of Bangladesh already moves its money.)

The significance is distribution. Bangladesh does not lack savings products; it lacks savings products that reach a garment worker in Savar or a shopkeeper in Rangpur without a branch, a form and an afternoon lost. bKash’s tens of millions of users are exactly the population formal banking has struggled to hold, and a deposit product living where their money already lives is financial inclusion by the shortest path.

Who benefits, and who pays? Semi-urban and rural savers benefit from formal, interest-bearing savings a few taps away; EBL gets deposits it could never reach through branches; bKash deepens its platform. The watch-item is consumer clarity — terms, early-withdrawal rules and deposit protection need to be as visible on a phone screen as on a branch poster.

The Dhaka view connects this to the week’s bigger arc: payments reform above, savings products here — the financial layer of Bangladesh’s digital economy is being assembled piece by piece. Every reliably banked household is also, eventually, a customer for every digital service this newsletter covers. Inclusion is infrastructure.

5. The promised deepfake panel’s notification did not appear

An accountability follow-up, because this newsletter made a note last week to check. On July 21, the government said a special panel of lawyers to pursue deepfakes, disinformation and cyberbullying was fully prepared, with the information adviser stating: “Only the formal government notification remains to be issued, which is expected within the next one or two days as the government has taken zero-tolerance policy against these offences.” As of Friday July 31 — ten days later — I could find no report of that notification in the state news agency or the major dailies, and nothing identifiable in the government printing office’s gazette listings for July. Two honest hedges: gazette titles are opaque, so absence from the listing is indicative rather than conclusive; and a notification issued quietly late in the week could surface in next week’s coverage. As always with this topic, the treatment here stays general and awareness-oriented — the point is the policy, not any individual case. (A plain-language note: a “gazette notification” is the official published order that gives a government decision legal force; until it issues, a promised body exists only in the briefing room.)

The significance is exactly what this newsletter wrote when the promise was made: “only the notification remains” is a sentence to hold lightly. The panel is the right idea — proactive enforcement against image-based abuse, which overwhelmingly targets women, is the missing piece of Bangladesh’s cyber-protection framework — which is precisely why its non-appearance matters more than a routine delay would.

Who benefits, and who pays? Nobody benefits from the gap between podium and gazette. The ones who pay are the people the panel was promised to protect, for whom every week without an enforcement body is another week of bearing the entire burden of reporting abuse themselves.

The Dhaka view is patient but unimpressed. Governments miss self-imposed deadlines constantly, and ten days is not a scandal. But this newsletter has tracked the pattern for months: the deepfake harm is named at press conferences and remains unnamed in law, capacity is promised and remains unconstituted. The ask is small and specific — issue the notification, staff the panel, publish its remit. Next Friday, this section will check again.

Wrapping up

Ten stories in seven days. Globally, the hyperscalers closed earnings season by showing the AI boom at full thrust: Microsoft’s Azure crossed $100 billion in annual revenue while accelerating to 43 per cent growth; Meta grew 28 per cent yet spent so aggressively its free cash flow nearly vanished, spooking its investors; Amazon’s AWS grew its fastest in 18 quarters with in-house chips crossing a $25 billion run rate. Against those ledgers, more than 1,300 frontier-lab researchers — including the chief scientists of nearly every major American lab — publicly asked Washington for the tools to deliberately pace frontier AI, and the EU AI Act reached the eve of real enforcement, fines included.

At home, the Prime Minister opened Bangladesh’s first national semiconductor summit, with eight roadmaps and genuine fiscal measures behind the design-first ambition; Bangladesh Bank produced its most coherent week in memory — a framework for PayPal-style cross-border payments, then a consolidated export rulebook with the first dedicated freelancer chapter; EBL and bKash put a bank deposit inside a mobile wallet; and the promised deepfake panel’s notification, due “within one or two days” of July 21, had still not appeared by Friday.

If there is one thread, it is that every story this week was about two ledgers. The industry’s financial ledger has never looked stronger — $100 billion clouds, $220 billion spending plans, backlogs measured in half-trillions. Its ledger of trust is the one its own researchers publicly worried over, the one Brussels will now audit with fines, and the one Dhaka keeps in its own small denominations: a circular that becomes a working payment, a roadmap that becomes a trained engineer, a briefing-room promise that does or does not become a gazette. Money compounds quickly and trust slowly, and the week’s quiet lesson — from Redmond to Motijheel — is that the second ledger is the one that decides what the first one was worth.

If something here caught your interest, share it with one person who does not follow tech. The conversation only spreads if we keep starting it.

See you next week. Stay well.